chunked/nltk_chunking/model_doc_imagegpt/chunk_22.txt

The authors showed that the
  features in the middle of the network are the most performant, and can be used as-is to train a linear model (such as
  a sklearn logistic regression model for example).